====== Perplexity statistics ======
Mean PPL(Q)                   :  25.241862 ±   0.238190
Mean PPL(base)                :  22.656280 ±   0.216110
Cor(ln(PPL(Q)), ln(PPL(base))):  96.17%
Mean ln(PPL(Q)/PPL(base))     :   0.108067 ±   0.002627
Mean PPL(Q)/PPL(base)         :   1.114122 ±   0.002927
Mean PPL(Q)-PPL(base)         :   2.585582 ±   0.066553

====== KL divergence statistics ======
Mean    KLD:   0.310675 ±   0.000978
Maximum KLD:  13.193255
99.9%   KLD:   3.235092
99.0%   KLD:   1.749219
99.0%   KLD:   1.749219
Median  KLD:   0.202102
10.0%   KLD:   0.005881
 5.0%   KLD:   0.001288
 1.0%   KLD:   0.000092
Minimum KLD:  -0.000012

====== Token probability statistics ======
Mean    Δp: -2.612 ± 0.035 %
Maximum Δp: 98.812%
99.9%   Δp: 61.048%
99.0%   Δp: 34.086%
95.0%   Δp: 14.732%
90.0%   Δp:  6.759%
75.0%   Δp:  0.325%
Median  Δp: -0.077%
25.0%   Δp: -3.968%
10.0%   Δp: -16.600%
 5.0%   Δp: -27.404%
 1.0%   Δp: -51.971%
 0.1%   Δp: -79.416%
Minimum Δp: -99.470%
RMS Δp    : 13.592 ± 0.056 %
Same top p: 74.228 ± 0.113 %